UAV-assisted deep learning to support results-based agri-environmental schemes: Facilitating Eco-Scheme 5 implementation in Germany

Abstract

Results-based agri-environmental schemes (AES) hold significant potential to promote biodiversity and ecosystem services within agricultural landscapes. However, a key obstacle to their widespread adoption is the practical challenge of verifying target species (result indicators) accurately and cost-effectively. This study presents a digital and automated approach to verify (result) indicators in grasslands to facilitate the implementation of Eco-Scheme 5, a results-based AES introduced in Germany. The presented approach employs a deep learning-based object detection framework to automatically detect indicator plant species in high-resolution RGB images acquired using unmanned aerial vehicles (UAVs). Additionally, the study explores whether incorporating ground-based imagery into the UAV training dataset could enhance model performance on UAV imagery, hypothesizing robust generalization across these image domains. The Baseline model, trained exclusively on UAV imagery, achieved an average precision (AP₅₀) of 74.0, with performance affected primarily by insufficient training data and class imbalance, particularly affecting species with fewer instances. In contrast, the Enhanced model, trained on UAV imagery enriched with ground-based data, achieved a significantly higher AP₅₀ of 94.2 on the UAV test dataset, demonstrating improved detection accuracy and robust cross-domain generalization. These findings validate the benefits of cross-domain training in improving model performance and emphasize the potential of UAV-integrated artificial intelligence for efficient biodiversity monitoring and supporting the implementation of results-based AES.